Exploiting second-order statistics improves statistical learning of vowels

Abstract

Unsupervised clustering algorithms were used to evaluate three models of statistical learning of minimal contrasts between English vowel pairs. The first two models employed only first-order statistics with assumptions of uniform [M1] or Gaussian [M2] distributions of vowels in an F1-F2 space. The third model [M3] employed second-order statistics by encoding covariance between F1 and F2. Acoustic measures of F1/F2 frequencies for 12 vowels spoken by 139 men, women, and children (Hillendrand et al. 1995) were used as input to the models. Effectiveness of each model was tested for each minimal-pair contrast across 100 simulations. Each simulation consisted of two centroids that adjusted on a trial-by-trial basis as 1000 F1/F2 pairs were input to the models. With addition of each pair, centroids were reallocated by a k-means algorithm, an unsupervised clustering algorithm that provides an optimal partition of the space into uniformly-sized convex cells. The first-order Gaussian model [M2] performed better th...